High-precision drug screening method, system and device based on hierarchical multi-modal representation learning and readable storage medium

Abstract

The application relates to the technical field of computational biology, and discloses a high-precision drug screening method and system based on hierarchical multi-modal representation learning, equipment and a readable storage medium, which comprise the following steps: performing feature extraction on an amino acid sequence of a target protein through a plurality of pre-trained protein language models; constructing multi-scale graph representation of the protein and ligand molecule graph representation based on the three-dimensional structure of the target protein; performing multi-modal feature learning on the multi-scale graph representation of the protein and the ligand molecule graph representation by using a message passing neural network; and constructing a binding affinity prediction model for drug screening based on the multi-modal features to complete a drug screening task. The complex interaction in the protein and ligand graph structure is effectively captured by using a multi-layer graph neural network, efficient coding of protein-ligand pair features is realized, surface, structure and ligand embedding vectors of the protein can be extracted, and the accuracy of protein-ligand interaction prediction is improved.

Description

Technical Field

[0001] This invention relates to the field of computational biology, specifically to a high-precision drug screening method, system, device, and readable storage medium based on hierarchical multimodal representation learning. Background Technology

[0002] Drug screening is a crucial step in new drug development, aiming to rapidly identify highly effective compounds that can bind to target proteins from a vast library of compounds. The binding affinity between proteins and ligands is a key indicator of drug potential. Traditional drug screening methods primarily rely on high-throughput experimental techniques, such as enzyme inhibition assays and surface plasmon resonance assays. While precise, these methods are time-consuming and costly, limiting their feasibility for large-scale application.

[0003] In recent years, computational virtual screening methods have attracted widespread attention. These methods rapidly assess the binding affinity of compounds by simulating the physical and chemical interactions between proteins and ligands, significantly improving the efficiency of drug screening. However, traditional computational models (such as molecular force fields and molecular docking) have limitations in handling the multi-scale characteristics of complex biological systems. For example, these methods often struggle to simultaneously capture the sequence, structure, and surface features of proteins, as well as the molecular properties of ligands.

[0004] With the development of artificial intelligence technology, deep learning methods have shown great potential in the field of drug screening. In particular, the application of Message Passing Neural Networks (MPNNs) in molecular graph representation learning has made the prediction of protein-ligand interactions more accurate. However, current models based on single-modal features have certain limitations in practical applications, such as the inability to effectively integrate multi-scale features of proteins or the neglect of global information from ligand molecular graphs. Therefore, developing a multimodal deep learning method that can integrate protein sequence, structure, surface features, and ligand molecular information has become one of the key challenges in improving the accuracy of drug screening.

[0005] In recent years, the rapid development of protein language models (PLMs) and multimodal representation learning techniques has provided new tools for this field. For example, pre-trained models such as Ankh, ESM-1b, and ProtTrans can extract deep semantic and evolutionary information from protein sequences, while triangular mesh-based protein surface modeling tools (such as MSMS) have laid the foundation for efficient representation of surface features. Meanwhile, the combination of the message passing mechanism of MPNNs and multilayer perceptron (MLP) regressors makes it possible to fuse multimodal features for binding affinity prediction. Summary of the Invention

[0006] In view of the above-mentioned problems, the present invention is proposed.

[0007] Therefore, the technical problem solved by this invention is the feature encoding problem in protein-ligand interaction prediction, especially the extraction of efficient and accurate embedding vectors from complex graph structure data.

[0008] To address the aforementioned technical problems, this invention provides the following technical solution: a high-precision drug screening method based on hierarchical multimodal representation learning, comprising:

[0009] The hierarchical multimodal representation learning method establishes a cascaded fusion strategy for protein multi-scale features.

[0010] Feature extraction of the amino acid sequence of the target protein is performed using multiple pre-trained protein language models;

[0011] Based on the three-dimensional structure of proteins, construct multi-scale graphical representations of proteins;

[0012] Based on the ligand molecular formula, the atomic-level structure diagram of the ligand molecule is constructed, which is referred to as the ligand molecule diagram representation;

[0013] Construct a message-passing neural network to achieve multimodal feature learning and representation of multiscale graphs of proteins and atomic structure graphs of ligands;

[0014] By integrating multimodal features, a computational model for predicting protein-ligand binding affinity is constructed to complete a high-throughput virtual drug screening task.

[0015] As a preferred embodiment of the high-precision drug screening method based on hierarchical multimodal representation learning described in this invention, the multiple pre-trained protein language models include Ankh, ESM-1b, and ProtTrans; feature extraction of protein amino acid sequences includes using the above three models to represent features of amino acid sequences of length L respectively; and fusing the three feature matrices to generate a hybrid embedding matrix M. B :

[0016]

[0017] Among them, M A M represents the amino acid sequence feature matrix extracted by the Ankh model; E M represents the amino acid sequence feature matrix extracted by the ESM-1b model. P M represents the amino acid sequence feature matrix extracted by the ProtTrans model; B This represents the hybrid embedding matrix obtained by fusing the three feature matrices. This indicates that the two feature matrices are concatenated "row-wise".

[0018] As a preferred embodiment of the high-precision drug screening method based on hierarchical multimodal representation learning described in this invention, the construction of multi-scale graph representation of proteins and ligand molecule graph representation includes constructing residue-level structural graph representation based on the three-dimensional structure of proteins;

[0019] The three-dimensional structure of a protein is represented as a residue-level diagram G. c =(V c E c );

[0020] Among them, V c ={v1,…,v L} and E c ={e1,…,e N} Corresponding to Figure G c A set of L nodes and N edges in a given context;

[0021] By determining whether the distance between the Cα atoms of any two amino acid residues is less than 1 / 3... To determine each edge in this residue-level graph;

[0022] Wherein, the i-th node v i Corresponding to any sequence S p The i-th residue in the sequence, 1≤i≤L;

[0023] node v i eigenvectors The formula is expressed as:

[0024]

[0025] When node v i There is K i When there are neighboring nodes, edge e ij Connect node v i and v j ;

[0026] Among them, v j ∈V c ,e ij ∈E c ,j≤K i <L;

[0027] Side e ij eigenvector ef ij The formula is defined as:

[0028] ef ij =[d ij ,θ ij ]

[0029] Where, d ij Indicates amino acid v i and v j The distance between them; θ ij Indicates v i and v j The angle between the normal vectors;

[0030] Use a triangular mesh generation tool to construct a surface map representation of proteins;

[0031] For a protein p, its surface is constructed as a triangular mesh using MSMS software;

[0032] By combining all the meshes, we obtain the surface map G. F =(V F E F );

[0033] Among them, V F E represents the set of vertices extracted from all meshes. F This represents the set of edges extracted from all grids;

[0034] Surface map G F Any vertex u in j ∈V F eigenvectors of surface maps Initialize it to the nearest structure graph G c Middle residue node v j The eigenvectors of are expressed by the formula:

[0035]

[0036] Edge e on the triangular mesh jk eigenvector ef jk The formula is expressed as:

[0037] ef jk =[d jk ,g jk ]

[0038] Where, d jk Represents vertex u j and u k The distance between them; g jk Let be a 6-dimensional vector representing the connection with edge e. jk The six geometric parameters between two adjacent grids include the dihedral angle θ between two adjacent faces, the interior angle (θ) between two opposite faces, and the angle between two adjacent faces. a ,θ b The ratio of the side length to the vertical height of each adjacent face (d / h) a ,d / h b), and the angle θ(n) between the normal vectors at the nodes that make up the edge. k ,n j );

[0039] Molecular diagram representations are constructed using the SMILES strings of ligands;

[0040] For a given 1D ligand SMILES string, use RDKit to obtain the molecular structure, represented as a 2D ligand diagram G. L =(V L E L );

[0041] Among them, V L E represents the set of vertices extracted from the ligand graph. L This represents the set of edges extracted from the ligand;

[0042] 2D ligand diagrams include atomic-type embeddings and physicochemical properties, linked by covalent bonds;

[0043] Atomic characteristics include atomic symbol, atomic degree, implicit valence, explicit valence, and whether it belongs to an aromatic ring;

[0044] The main characteristics of a bond include the bond type, whether it is a conjugate bond, and whether it belongs to a ring.

[0045] As a preferred embodiment of the high-precision drug screening method based on hierarchical multimodal representation learning described in this invention, the multimodal feature learning includes inputting the protein's structural map, surface map, and ligand molecule map, and performing multi-level message passing;

[0046] Construct an N-layer message-passing neural network, starting from the structure graph G c Surface diagram G F and ligand diagram G L Different protein-ligand pair features are generated in the process;

[0047] An N-layer message-passing neural network has its input defined as graph G = (V, E);

[0048] In graph G, the i-th node v i The set of neighbor nodes of ∈V is represented as NP(v i The formula for initializing the message of this node is expressed as:

[0049]

[0050] Where "0" represents the initial state of the network model at time t=0. Represents node v i The message vector at layer 0, Represents node vi eigenvectors;

[0051] Each node v in the message-passing neural network i The formula for message passing is expressed as:

[0052]

[0053] in, Represents node v i The message vector at layer k, where τ represents the activation function, U1 represents the weight matrix, U2 represents the weight matrix, and NP(v i ) represents node v i The set of neighboring nodes, Represents node v i and its neighbor node v j The attention weights in the (k-1)th layer, where Q represents the weight matrix. Represents node v i eigenvectors and edge e ij eigenvector ef ij To concatenate the messages, N represents the total number of message passing layers, from layer 1 to layer N; Represents node v i The initial feature vector, ef ij Representing edge e ij The feature vector, containing the connection node v i and v j Relationship information; a represents scalar coefficients, W represents the weight matrix, ev represents the final feature embedding vector, W (1) W (2) W (3) Let ev represent the learnable weight matrix. f An embedding vector representing a protein surface feature, ev c An embedding vector representing protein structural features, ev l An embedding vector representing the characteristics of a ligand molecule.

[0054] As a preferred embodiment of the high-precision drug screening method based on hierarchical multimodal representation learning described in this invention, the multimodal feature learning further includes introducing an attention mechanism to dynamically adjust the information weights of different neighboring nodes during message passing, expressed by the following formula:

[0055]

[0056] in, This indicates that at level k-1, node v i Its neighbor node v jAttention weights between them; exp represents the natural exponential function, and LReLU represents the LeakyReLU activation function. Represents node v i Its neighbor node v j The attention pre-value at layer k-1, NP(v) i ) represents node v i The set of neighboring nodes, Represents node v i With node v q The pre-attention value at layer k-1; Represents node v i and its neighbor node v j In the attention pre-values ​​of the (k-1)th layer, 'a' represents a scalar coefficient, and 'W' represents the learnable weight matrix. Represents node v i The message vector at layer k-1, Represents node v j The message vector at the (k-1)th layer; || represents the concatenation operation;

[0057] After N layers of message passing, the protein's structure embedding vector, surface embedding vector, and ligand embedding vector are extracted, as expressed by the following formula:

[0058]

[0059] Where ev represents the final feature embedding vector, NP(v i ) represents node v i The set of neighboring nodes, W (1) This represents the learnable weight matrix. Represents node v j The message vector at layer N, W (2) Let ef represent the learnable weight matrix. ij Representing edge e ij eigenvectors, W (3) This represents the learnable weight matrix. Represents node v i In the message vector at layer N, V ​​represents the set of all nodes in graph G;

[0060] For any protein-ligand pair, the MPNN module outputs three embedding vectors of length 200; the three embedding vectors of length 200 include a surface vector, a structure vector, and a ligand vector.

[0061] As a preferred embodiment of the high-precision drug screening method based on hierarchical multimodal representation learning described in this invention, the construction of a binding affinity prediction model for drug screening includes concatenating surface vectors, structure vectors, and ligand vectors to generate a unified multimodal feature representation of protein-ligand pairs.

[0062] As a preferred embodiment of the high-precision drug screening method based on hierarchical multimodal representation learning described in this invention, the construction of the binding affinity prediction model for drug screening further includes inputting a unified multimodal feature representation of protein-ligand pairs into a multilayer perceptron regression model to predict the binding affinity value between the protein and the ligand.

[0063] The multilayer perceptron regression model is trained by minimizing the mean squared error, as expressed by the formula:

[0064]

[0065] in, Represents the predicted value, y i represents the true value, and MSE represents the multilayer perceptron regression model.

[0066] A high-precision drug screening system based on hierarchical multimodal representation learning, wherein:

[0067] The feature extraction module extracts features from the amino acid sequence of the target protein using multiple pre-trained protein language models.

[0068] The graph representation module constructs multi-scale graph representations of proteins and ligand molecule graph representations based on the three-dimensional structure of the target protein.

[0069] The multimodal feature module uses a message-passing neural network to learn multimodal features by constructing multi-scale graph representations of proteins and ligand molecular graph representations.

[0070] The screening module, based on multimodal features, constructs a binding affinity prediction model for drug screening to complete the drug screening task.

[0071] A computer device includes: a memory and a processor; the memory stores a computer program, characterized in that: when the processor executes the computer program, it implements the steps of the method described in any one of the present invention.

[0072] A computer-readable storage medium having a computer program stored thereon, characterized in that: when the computer program is executed by a processor, it implements the steps of the method described in any one of the present invention.

[0073] The beneficial effects of this invention are as follows: The high-precision drug screening method based on hierarchical multimodal representation learning provided by this invention effectively captures the complex interaction relationships in the protein-ligand graph structure by using a multilayer graph neural network (MPNN), achieving efficient encoding of protein-ligand pair features. This method can extract the surface, structure, and ligand embedding vectors of proteins, improving the accuracy and efficiency of protein-ligand interaction prediction. Simultaneously, by introducing an attention mechanism and graph convolutional networks, the information transmission process is optimized, giving the model stronger robustness and generalization ability when processing large-scale molecular graphs. Attached Figure Description

[0074] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0075] Figure 1 This is an overall flowchart of a high-precision drug screening method based on hierarchical multimodal representation learning, provided in the first embodiment of the present invention.

[0076] Figure 2 This is an overall framework diagram of HMRLBA, a high-precision drug screening method based on hierarchical multimodal representation learning, provided in the fourth embodiment of the present invention.

[0077] Figure 3 The ablation experiment results of HMRLBA, a high-precision drug screening method based on hierarchical multimodal representation learning, provided in the fourth embodiment of the present invention;

[0078] Figure 4 Contribution analysis of different language models in HMRLBA, a high-precision drug screening method based on hierarchical multimodal representation learning, provided in the fourth embodiment of the present invention;

[0079] Figure 5 The experimental results comparing HMRLBA, a high-precision drug screening method based on hierarchical multimodal representation learning, with mainstream methods provided in the fourth embodiment of the present invention;

[0080] Figure 6 The fourth embodiment of the present invention provides a high-precision drug screening method based on hierarchical multimodal representation learning (HMRLBA) and its prediction results for difficult samples in the test set, combined with visualization.

[0081] Figure 7The flowchart illustrates the application of HMRLBA, a high-precision drug screening method based on hierarchical multimodal representation learning, in EGFR virtual drug screening, as provided in the fourth embodiment of the present invention. Detailed Implementation

[0082] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0083] Example 1, referring to Figure 1 As an embodiment of the present invention, a high-precision drug screening method based on hierarchical multimodal representation learning is provided, comprising:

[0084] S1: Feature extraction of the amino acid sequence of the target protein is performed using multiple pre-trained protein language models.

[0085] For the PDB file of protein p, first use BioPython to extract its amino acid sequence S. p The sequence length is L. In the protein language model system, the protein sequence S can be represented by three models: Ankh, ESM-1b, and ProtTrans. p Represented as three embedding matrices M A M E , and M P The sizes of these three matrices are L×1536, L×1280, and L×1024, respectively.

[0086] (2) Concatenate the embedded feature matrices to generate a hybrid embedding matrix.

[0087] The three embedding matrices mentioned above are merged into a single hybrid embedding matrix M. B :

[0088]

[0089] Furthermore, by concatenating embedding feature matrices from different pre-trained protein language models (Ankh, ESM-1b, and ProtTrans), the advantages of multiple models are integrated, thereby fully utilizing the different understandings and expression capabilities of different models regarding amino acid sequence features. This hybrid embedding matrix enables a more comprehensive capture of multidimensional information about protein sequences, improving the accuracy and robustness of downstream tasks (such as protein structure prediction and functional annotation). Simultaneously, this method integrates the learning capabilities of different models, enhances the diversity of feature expression, and improves the model's ability to identify protein structure and function.

[0090] S2: Based on the three-dimensional structure of the target protein, construct multi-scale graphical representations of the protein and ligand molecular representations.

[0091] Constructing multi-scale graphical representations of proteins can comprehensively capture the internal and external structural features and geometric relationships of proteins, which plays a crucial role in understanding the biological functions of proteins and their interactions with ligands. Residue-level structural representations can visually display the spatial relationships and local geometric properties between amino acids, while protein surface representations constructed using triangular mesh generation tools characterize the spatial features of the protein's exterior. This multi-scale representation provides a solid foundation for subsequent feature extraction and learning. Furthermore, combining the multi-scale features of proteins with the molecular graphical representations of ligands can comprehensively reflect the multimodal characteristics of the protein-ligand system, laying an important foundation for multimodal feature fusion in affinity prediction.

[0092] Residue-level structural diagrams constructed based on the three-dimensional structure of proteins include:

[0093] The three-dimensional structure of a protein provides internal details, including distances, angles, torsion angles, and other higher-order terms. Here, we represent the protein structure as a residue-level diagram G. c =(V c E c ), where V c ={v1,…,v L} and E c ={e1,…,e N} corresponds to a set of L nodes and N edges, respectively. Graph G c Each edge in the algorithm is determined by whether the distance between the Cα atoms of any two amino acid residues is less than 1 / 3. It was determined that.

[0094] The i-th node v i Corresponding to sequence S p The i-th residue in the matrix, where 1 ≤ i ≤ L. Therefore, node v i eigenvectors Initialize to:

[0095]

[0096] If node v i has K i neighbor nodes, then edge e ij connects node v i and v j (where v j ∈V c , e ij ∈E c , j≤K i <L). The edge e ij can be initialized with its feature vector ef ij as follows:

[0097] ef ij =[d ij , θ ij (3)

[0098] where d ij represents the distance between amino acid v i and v j ; in addition, θ ij represents the included angle between the normal vectors of these two nodes.

[0099] Constructing a surface graph representation of a protein using a triangular mesh generation tool includes:

[0100] The surface representation shows the external details of the protein. For a protein p, its surface is constructed as a triangular mesh using MSMS software. Combining all meshes, we can obtain a surface graph G F =(V F , E F ). Where V F and E F are respectively the set of vertices and edges extracted from all meshes. For any vertex u F in G j ∈V F , its feature vector is initialized to the feature vector of the residue node v c that is closest to it in the structure graph G j :

[0101]

[0102] The feature vector ef of the edge e on the triangular mesh jk can be initialized as follows: jk

[0103] ef jk =[d​jk ,g jk (5)

[0104] Where, d jk It is vertex u j and u k The distance between them; g jk Defined as a 6-dimensional vector, representing the connection with edge e. jk Six geometric parameters between two adjacent grids that are associated.

[0105] Constructing molecular diagram representations using the SMILES strings of ligands includes:

[0106] For a given 1D ligand SMILES string, HMRLBA uses RDKit to obtain the molecular structure and further represents it as a 2D ligand graph G. L =(V L E L This diagram contains embedded and physicochemical properties based on atom type, connected by covalent bonds. Atom characteristics include atom symbol, atomic degree, implicit valence, explicit valence, and whether it belongs to an aromatic ring; while bond characteristics mainly include bond type, whether it is a conjugated bond, and whether it belongs to a ring.

[0107] Furthermore, by constructing multi-scale graph representations of proteins and molecular graph representations of ligands, we can comprehensively capture the internal and external structural features of proteins and their geometric relationships with ligands, providing rich structural information for subsequent feature extraction and learning. Residue-level structural graphs can reveal the spatial relationships between amino acids, while surface graphs provide the external geometric features of proteins, thus providing a more accurate foundation for understanding protein-ligand interactions. Molecular graph representations of ligands, through modeling molecular structure and physicochemical properties, comprehensively reflect the chemical characteristics of ligands. Combining these two multimodal features provides crucial multi-dimensional information support for subsequent protein-ligand affinity prediction and other biological tasks.

[0108] S3: Use message-passing neural networks to construct multi-scale graph representations of proteins and ligand molecular graph representations, and perform multimodal feature learning.

[0109] Input protein structure diagrams, surface diagrams, and ligand molecule diagrams to perform multi-level message passing.

[0110] We constructed a message-passing neural network (MPNN) to extract data from the above structure graph G. c Surface diagram G F and ligand diagram G L Different protein-ligand pair feature codes are generated. As a spatially based convolutional network, MPNN can model the complex interaction relationships between nodes and edges in a graph.

[0111] For an input graph G = (V, E), the processing steps of an N-layer MPNN model include the following. For node v i ∈V, its neighbor node set is represented as NP(v i The message initialization at this node is performed according to the following formula:

[0112]

[0113] Next, message passing from layer (k-1) to layer k is calculated using the following formula:

[0114]

[0115] Where 1≤k≤N. τ is the τ activation function, and U1, U2, and Q are weight matrices. In formula (7), from node v... i to v j Attention score α ij The definition is as follows:

[0116]

[0117] The coefficient β in formula (8) ij It can be calculated using formula (9):

[0118]

[0119] Extracting the protein's structure embedding vector, surface embedding vector, and ligand embedding vector includes:

[0120] Message aggregation of graph-encoded vectors is implemented in the following way:

[0121]

[0122] Where the operation ⊙ represents the Hadamard product, W (1) W (2) and W (3) It is a learnable weight matrix. For any protein-ligand pair, the MPNN module outputs three embedding vectors of length 200, namely the surface vector (ev...). f ), structure vector (ev) c ) and ligand vector (ev l ).

[0123] Furthermore, by introducing a multi-scale graph representation method based on message-passing neural networks (MPNN), multimodal feature learning is effectively performed on protein structure graphs, surface graphs, and ligand molecular graphs. MPNN can model the complex interactions between nodes and edges in the graph, thereby extracting multi-level features of protein-ligand pairs. This method yields embedding vectors for surfaces, structures, and ligands, thus improving the accuracy and robustness of protein-ligand interaction predictions. This design provides a more comprehensive and accurate feature representation for a deeper understanding of protein-ligand interactions, contributing to efficient drug screening and development in the biomedical field.

[0124] S4: Based on multimodal features, construct a binding affinity prediction model for drug screening to complete the drug screening task.

[0125] Generate a unified multimodal feature representation

[0126] The appeal surface vector (ev) f ), structure vector (ev) c ) and ligand vector (ev l The protein-ligand pairs are spliced ​​together to generate a unified multimodal feature representation.

[0127] The binding affinity values ​​between proteins and ligands are output using a multilayer perceptron regression model for drug screening tasks.

[0128] The multimodal feature representations of the protein-ligand pairs described above are input into a multilayer perceptron (MLP) regression model to predict the binding affinity between the protein and the ligand. The hidden layer of the MLP model contains 512 neurons, and its output represents the binding strength between the protein and the corresponding ligand, which can be directly used for drug screening tasks. As a regression model, the MLP is trained by minimizing the mean squared error (MSE), as shown in the following formula:

[0129]

[0130] in, and y i These represent the predicted value and the actual value, respectively.

[0131] Furthermore, by employing a unified multimodal feature representation, the surface, structure, and ligand embedding vectors of protein-ligand pairs are combined to form a more comprehensive feature description, improving the model's accuracy and generalization ability in drug screening tasks. Predicting binding affinity using a multilayer perceptron (MLP) regression model can effectively quantify the interaction strength between proteins and ligands, providing accurate affinity scores for drug screening, thereby accelerating the screening process of potential drug candidates and reducing the cost and time of experimental screening.

[0132] Example 2, an embodiment of the present invention, provides a high-precision drug screening system based on hierarchical multimodal representation learning, comprising:

[0133] The feature extraction module extracts features from the amino acid sequence of the target protein using multiple pre-trained protein language models.

[0134] The graph representation module constructs multi-scale graph representations of proteins and ligand molecular graph representations based on the three-dimensional structure of the target protein.

[0135] The multimodal feature module uses a message-passing neural network to learn multimodal features by constructing multi-scale graph representations of proteins and ligand molecular graph representations.

[0136] The screening module, based on multimodal features, constructs a binding affinity prediction model for drug screening to complete the drug screening task.

[0137] Example 3, an embodiment of the present invention, differs from the previous two embodiments in that:

[0138] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0139] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0140] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0141] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0142] Example 4 presents a high-precision drug screening method based on hierarchical multimodal representation learning, comprising the following steps: First, protein amino acid sequences are embedded using multiple pre-trained protein language models to generate a hybrid embedding matrix, preserving evolutionary information from different data sources; then, a multi-scale graph representation of the protein (including a residue-level structure graph based on three-dimensional structure and a surface graph generated through triangular meshes) and a molecular graph representation of the ligand are constructed to capture the internal and external features of the protein and the chemical information of the ligand; next, a message passing neural network (MPNN) is used to learn multimodal features from the protein's structure graph, surface graph, and ligand graph, extracting the embedding vectors of the protein and ligand; finally, the multimodal features are fused and input into a multilayer perceptron (MLP) regression model, outputting the binding affinity value between the protein and ligand for drug screening tasks. By combining the sequence, structure, and surface features of the protein with the multimodal information of the ligand molecule, this method significantly improves the accuracy and efficiency of drug screening and has broad application prospects.

[0143] Figure 1This is a diagram illustrating the overall framework of the high-precision drug screening method HMRLBA described in this invention. HMRLBA comprises four main steps: protein amino acid sequence feature extraction, construction of protein multi-scale graph representation and ligand molecular graph representation, multimodal feature learning, and construction of a binding affinity prediction model. The diagram details the generation of protein sequence features based on a pre-trained language model, the construction of protein surface graphs, structural graphs, and ligand molecular graphs, the multimodal feature learning process based on a message-passing neural network, and how to fuse multimodal features and input them into a multilayer perceptron regression model to ultimately achieve high-precision prediction of binding affinity.

[0144] Figure 2 The ablation experiments of HMRLBA, the method described in this invention, are presented, demonstrating the ablation results of the HMRLBA method under different module removal conditions. The ablation experiments evaluated the impact of protein structure map representation, surface map representation, and the attention mechanism in MPNN on the overall model performance.

[0145] Figure 3 This paper analyzes the contribution of different language models in HMRLBA, the method described in this invention, and demonstrates the contribution of pre-trained language models (Ankh, ESM-1b, and ProtTrans) used in the HMRLBA method to protein sequence feature extraction. Through experimental comparison, the figure shows the impact of features generated by each model on binding affinity prediction performance, reflecting the advantages of multilingual model fusion and the complementarity of each model in terms of feature expression capabilities.

[0146] Figure 4 The experimental results comparing the HMRLBA method described in this invention with mainstream methods are presented. These results show that the HMRLBA method is compared with mainstream methods (including sequence-based methods DeepDTA, MGraphDTA, and PSICHC; structure-based methods IEConv; surface-based methods MaSIF; and multi-scale models HoloProt and HaPPy). The results are based on root mean square error (RMSE), mean absolute error (MAE), Pearson correlation coefficient, Spearman correlation coefficient, and coefficient of determination (R²). 2 Five performance metrics were evaluated. The results show that HMRLBA outperforms other methods on all metrics, exhibiting the lowest errors in RMSE and MAE, indicating higher prediction accuracy; it also achieves the highest values ​​in Pearson and Spearman correlation coefficients, reflecting its superior performance in both linear and nonlinear relationship prediction; simultaneously, its R... 2 The values ​​are significantly higher than other methods, demonstrating stronger interpretability. These results validate the superior performance and reliability of the HMRLBA method in drug screening tasks.

[0147] Figure 5 This paper presents the prediction results and binding visualization of the HMRLBA method described in this invention for difficult samples in the test set. The binding sites of four "difficult samples" (complexes 1Q7A, 4QGD, 4TKJ, and 5HZ6) are visualized using PyMOL software, along with the actual binding affinity and predicted affinity of the HMRLBA model. As can be seen from the figure, these ligands bind near the protein surface, and the prediction error between the HMRLBA-predicted affinity and the actual affinity is minimal, achieving very satisfactory results.

[0148] Figure 6 This is a flowchart illustrating the application of the HMRLBA method described in this invention in virtual drug screening for EGFR. We first retrieved the SMILES structures of 2,616 FDA-approved small molecule drugs from the DrugBank database, including 13 EGFR inhibitors. Simultaneously, we obtained the PDB structure files of EGFR from the PDB database. These data were then input into a trained HMRLBA model to obtain predicted binding affinity values. Table 1 shows the parameter settings for the HMRLBA method described in this invention.

[0149] Table 1 Parameter Setting Table

[0150]

[0151]

[0152] Table 2 presents the ablation experiment results of HMRLBA, the method described in this invention. Specifically, we analyzed the contribution of each component to model performance by removing the protein surface representation (woSurface), structural representation (woStructure), and attention mechanism (woAttention) from the MPNN module, respectively. The rest of the model architecture remained unchanged except for the ablated components. The experimental results show that the HMRLBA model significantly outperforms the three ablation models, with the removal of the attention mechanism having the greatest impact on performance, followed by the removal of the surface or structural representation. This is because the attention mechanism enhances the weight allocation ability during message passing, the structural representation provides information on the three-dimensional conformation of the protein, and the surface representation captures the shape and geometric distribution of the protein surface. Therefore, we conclude that protein multimodal feature learning and the attention mechanism can improve the accuracy of the model in predicting protein-ligand binding affinity.

[0153] Table 2 Ablation Experiment Results

[0154] woAttention 1.512±0.011 1.221±0.009 0.478±0.007 0.475±0.006 0.196±0.014 woStructure 1.495±0.011 1.198±0.008 0.519±0.006 0.516±0.006 0.207±0.012 woSurface 1.473±0.009 1.182±0.007 0.492±0.005 0.487±0.006 0.234±0.011 HMRLBA 1.428±0.007 1.146±0.005 0.544±0.004 0.537±0.004 0.280±0.008

[0155] Table 3 presents the contribution analysis of different language models in HMRLBA as described in this invention, investigating the contributions of three different protein language models used in HMRLBA. We compared the experimental performance of seven feature embeddings in the sequence embedding module, including three independent embeddings from ProtTrans, ESM-1b, and Ankh, and four hybrid embeddings created by concatenating these embeddings: ProtTrans+ESM-1b (PE), ProtTrans+Ankh (PA), ESM-1b+Ankh (EA), and a combination of the three (PEA = HMRLBA). The experimental results compared the performance of the seven feature embeddings, showing that PEA significantly outperformed the other embedding methods. Among the three independent embeddings, ESM-1b performed best. Furthermore, when ESM-1b was combined with PA, the model performance was significantly improved. Based on these findings, two conclusions can be drawn. First, the complementarity among the three protein language models pre-trained on different protein sequence databases significantly improved the model's accuracy in predicting protein-ligand binding affinity. Second, ESM-1b played a crucial role in all three protein language models.

[0156] Table 3. Contribution Analysis of Feature Embedding Achieved by Different Large Language Models

[0157]

[0158] Table 4 shows the comparative experimental results of HMRLBA, the method described in this invention, with mainstream methods. The results show that HMRLBA is significantly superior to other methods. Furthermore, HMRLBA exhibits comparable or smaller standard deviations across all evaluation metrics compared to other methods, indicating lower volatility and higher model stability in its test results.

[0159] Table 4 Comparison of experimental results

[0160] DeepDTA 1.774±0.019 1.433±0.012 0.463±0.007 0.457±0.008 0.086±0.017 MGraphDTA 1.692±0.012 1.354±0.008 0.353±0.009 0.336±0.010 0.107±0.012 PSICHIC 1.539±0.014 1.235±0.009 0.504±0.011 0.498±0.010 0.166±0.013 IEConv 1.569±0.006 1.257±0.005 0.426±0.016 0.434±0.014 0.158±0.008 MaSIF 1.487±0.009 1.195±0.006 0.476±0.008 0.465±0.007 0.221±0.007 Holoprot 1.524±0.011 1.228±0.008 0.485±0.007 0.473±0.006 0.182±0.013 HaPPy 1.469±0.013 1.180±0.011 0.526±0.009 0.519±0.008 0.242±0.012 HMRLBA 1.428±0.007 1.146±0.005 0.544±0.004 0.537±0.004 0.280±0.008

[0161] Table 5 shows the prediction results of the HMRLBA method described in this invention for "hard samples" in the test set. In the field of protein-ligand affinity prediction, protein sequence homology is one of the key factors affecting the generalization ability of the model. To accurately evaluate the model's performance on low homology samples, this study uses the NW-align algorithm to calculate protein homology and introduces the concept of "hard samples," defined as samples in the test set whose protein homology is less than 30% compared to the training set. In this experiment, we systematically compared the performance of four methods—HMRLBA, HaPPy, MaSIF, and HoloProt—on 21 hard samples.

[0162] Table 5 Prediction Results for Difficult Samples

[0163]

[0164]

[0165] Table 6 presents the experimental results of the HMRLBA method and comparative methods described in this invention in virtual drug screening for EGFR. After confirming the effectiveness and stability of the HMRLBA model, we applied it to virtual drug screening for epidermal growth factor receptor (EGFR). The experimental results show the ranking of different methods for predicting affinity of EGFR small molecule inhibitors.

[0166] Table 6. Virtual Screening Results of EGFR Inhibitors

[0167]

[0168] In summary, this invention proposes a high-precision drug screening method based on hierarchical multimodal representation learning, which is widely applicable to the fields of drug screening and drug discovery. The method has a fixed input data format, is easy to obtain from existing databases, and has a simple algorithm implementation with low computational resource requirements, thus possessing extremely high potential for widespread application.

[0169] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A high-precision drug screening method based on hierarchical multimodal representation learning, characterized in that, include: Hierarchical multimodal representation learning is used to establish a cascaded fusion strategy for protein multi-scale features. Feature extraction of the amino acid sequence of the target protein is performed using multiple pre-trained protein language models; Constructing multi-scale graphical representations of proteins based on their three-dimensional structures. Based on the ligand molecular formula, an atomic-level structure diagram of the ligand molecule is constructed, which is referred to as the ligand molecule diagram representation. Construct a message-passing neural network to achieve multimodal feature learning and representation of multiscale graphs of proteins and atomic structure graphs of ligands; By integrating multimodal features, a computational model for predicting protein-ligand binding affinity was constructed to complete a high-throughput virtual drug screening task. The pre-trained protein language models include Ankh, ESM-1b, and ProtTrans; feature extraction of protein amino acid sequences includes using the above three models to represent features of amino acid sequences of length L; The three feature matrices are fused to generate a hybrid embedding matrix. : in, This represents the amino acid sequence feature matrix extracted by the Ankh model; This represents the amino acid sequence feature matrix extracted by the ESM-1b model. This represents the amino acid sequence feature matrix extracted by the ProtTrans model; This represents the hybrid embedding matrix obtained by fusing the three feature matrices. This indicates that the two feature matrices are concatenated "row-wise"; The construction of multi-scale graph representations and ligand graph representations of proteins includes constructing residue-level structural graph representations based on the three-dimensional structure of proteins; Representing the three-dimensional structure of a protein as a residue-level diagram ; in, and Corresponding figures In Each node and The set of edges; By determining any two amino acid residues Whether the distance between atoms is less than 12 Å is used to determine each edge in the residue hierarchy diagram; Among them, the Nodes Corresponding to any sequence The first in One residue, ; node eigenvectors The formula is expressed as: When node have When there are neighboring nodes, the edge Connecting nodes and ; in, , , ; side eigenvectors The formula is defined as: in, Indicates amino acids and The distance between them; express and The angle between the normal vectors; Use a triangular mesh generation tool to construct a surface map representation of proteins; For a protein Its surface is constructed using a triangular mesh via MSMS software; By combining all the meshes, a surface map is obtained. ; in, This represents the set of vertices extracted from all meshes. This represents the set of edges extracted from all grids; Surface diagram any vertex in eigenvectors of surface maps Initialize it to the nearest structure graph Middle residue nodes The eigenvectors of are expressed by the formula: Edges on the triangular mesh eigenvectors The formula is expressed as: in, Represents vertices and The distance between them; Let represent a 6-dimensional vector, indicating the relationship between the edge and the vector. The six geometric parameters between two adjacent grids are related, including the dihedral angle between two adjacent faces. The inner angle opposite the side ( , The ratio of the side length to the vertical height of each adjacent face. , The angle between the normal vectors at the nodes that make up the edge and the edge itself. ; Molecular diagram representations are constructed using the SMILES strings of ligands; For a given 1D ligand SMILES string, use RDKit to obtain the molecular structure, which is represented as a 2D ligand diagram. ; in, This represents the set of vertices extracted from the ligand graph. This represents the set of edges extracted from the ligand; 2D ligand diagrams include atomic-type embeddings and physicochemical properties, linked by covalent bonds; Atomic characteristics include atomic symbol, atomic degree, implicit valence, explicit valence, and whether it belongs to an aromatic ring; The key characteristics of a bond mainly include the bond type, whether it is a conjugate bond, and whether it belongs to a ring.

2. The high-precision drug screening method based on hierarchical multimodal representation learning as described in claim 1, characterized in that: multimodal feature learning includes inputting the structural map, surface map and ligand molecule map of the protein, and performing multi-level message passing; Build Layered message-passing neural network, from the structure diagram Surface diagram and ligand diagram Different protein-ligand pair features are generated in the process; one Layered message-passing neural networks, whose input is defined as a graph ; picture The i-th node The set of neighbor nodes is represented as The formula for initializing messages for this node is expressed as: in, "0" indicates the initial state of the network model at time t=0. Represents a node In the message vector at layer 0, Represents a node eigenvectors; Each node in the message-passing neural network The formula for message passing is expressed as: in, Represents a node In the The message vector of the layer, This represents the activation function. Represents the weight matrix. Represents the weight matrix. Represents a node The set of neighboring nodes, Represents a node and its neighboring nodes In the Attention weights of layers Represents the weight matrix. Represents a node eigenvectors and edge eigenvectors To splice, This indicates the total number of message passing layers, from layer 1 to layer 2. layer; Represents a node The initial feature vector, Representing an edge The feature vectors, containing the connection nodes and Relationship information; Represents scalar coefficients. Represents the weight matrix. This represents the final feature embedding vector. This represents the learnable weight matrix. An embedding vector representing a protein surface feature. An embedding vector representing the structural features of a protein. An embedding vector representing the characteristics of a ligand molecule.

3. The high-precision drug screening method based on hierarchical multimodal representation learning as described in claim 2, characterized in that: Multimodal feature learning also includes introducing an attention mechanism during message passing to dynamically adjust the information weights of different neighbor nodes, expressed by the formula: in, Indicates the first Layers, nodes Its neighboring nodes Attention weights between them; LReLU represents the natural exponential function, and LReLU represents the Leaky ReLU activation function. Represents a node Its neighboring nodes In the Attention pre-value of the layer Represents a node The set of neighboring nodes, Represents a node With nodes In the The pre-attention value of the layer; Represents a node and its neighboring nodes In the Attention pre-value of the layer Represents scalar coefficients. This represents the learnable weight matrix. Represents a node In the The message vector of the layer, Represents a node In the The message vector of the layer; || represents the concatenation operation; go through After layer message passing, the protein's structure embedding vector, surface embedding vector, and ligand embedding vector are extracted, as expressed by the formula: in, This represents the final feature embedding vector. Represents a node The set of neighboring nodes, This represents the learnable weight matrix. Represents a node In the The message vector of the layer, This represents the learnable weight matrix. Representing an edge eigenvectors, This represents the learnable weight matrix. Represents a node In the The message vector of the layer, Representation diagram The set of all nodes in the set; For any protein-ligand pair, the MPNN module outputs three embedding vectors of length 200; the three embedding vectors of length 200 include a surface vector, a structure vector, and a ligand vector.

4. The high-precision drug screening method based on hierarchical multimodal representation learning as described in claim 3, characterized in that: Constructing a binding affinity prediction model for drug screening involves concatenating surface vectors, structure vectors, and ligand vectors to generate a unified multimodal feature representation of protein-ligand pairs.

5. The high-precision drug screening method based on hierarchical multimodal representation learning as described in claim 4, characterized in that: The construction of a binding affinity prediction model for drug screening also includes inputting a unified multimodal feature representation of protein-ligand pairs into a multilayer perceptron regression model to predict the binding affinity value between the protein and the ligand. The multilayer perceptron regression model is trained by minimizing the mean squared error, as expressed by the formula: in, Indicates the predicted value. Represents the actual value. This represents a multilayer perceptron regression model.

6. A high-precision drug screening system based on hierarchical multimodal representation learning employing the method described above, wherein the high-precision drug screening method based on hierarchical multimodal representation learning as described in any one of claims 1 to 5 is characterized in that: The feature extraction module extracts features from the amino acid sequence of the target protein using multiple pre-trained protein language models. The graph representation module constructs multi-scale graph representations of proteins and ligand molecule graph representations based on the three-dimensional structure of the target protein. The multimodal feature module uses a message-passing neural network to learn multimodal features by constructing multi-scale graph representations of proteins and ligand molecular graph representations. The screening module, based on multimodal features, constructs a binding affinity prediction model for drug screening to complete the drug screening task.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the high-precision drug screening method based on hierarchical multimodal representation learning as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the high-precision drug screening method based on hierarchical multimodal representation learning as described in any one of claims 1 to 5.

Images